Markov chain Monte Carlo

Transcription

1 Markov chain Monte Carlo Feng Li School of Statistics and Mathematics Central University of Finance and Economics Revised on April 24, 2017

2 Today we are going to learn... 1 Markov Chains 2 Metropolis Algorithm 3 Metropolis-Hastings 4 Multiple variables Feng Li (SAM.CUFE.EDU.CN) Statistical Computing 2 / 66

3 Markov Chains The goal of today s lecture is to learn about the Metropolis Hastings algorithm The Metropolis Hastings algorithm allows us to simulate from any distribution as long as we have the kernel of the density of the distribution. To understand the Metropolis Hastings algorithm, we must learn a little bit about Markov chains Feng Li (SAM.CUFE.EDU.CN) Statistical Computing 3 / 66

4 Basic Probability Rules Law of conditional probability Pr(A = a, B = b) = Pr(A = a B = b)pr(b = b) (1) More general conditional probability Pr(A = a, B = b C = c) = Pr(A = a B = b, C = c)ˆ Pr(B = b C = c) (2) Feng Li (SAM.CUFE.EDU.CN) Statistical Computing 4 / 66

5 Basic Probability Rules Marginalizing (for a discrete variable) Pr(A = a) = ÿ b Pr(A = a, B = b) (3) More general Pr(A = a C = c) = ÿ b Pr(A = a, B = b C = c) (4) Feng Li (SAM.CUFE.EDU.CN) Statistical Computing 5 / 66

6 Independence Two variables are independent if Pr(A = a, B = b) = Pr(A = a)pr(b = b (5) Dividing both sides by Pr(B=b) gives Pr(A = a B = b) = Pr(A = b (6) Feng Li (SAM.CUFE.EDU.CN) Statistical Computing 6 / 66

7 Conditional Independence Two variables A and B are Conditionally Independent if Pr(A = a, B = b C = c) = Pr(A = a C = c)ˆ Pr(B = b C = b, c (7) Dividing both sides by Pr(B = b C = c) gives Pr(A = a B = b, C = c) = Pr(A = a C = b, c (8) Feng Li (SAM.CUFE.EDU.CN) Statistical Computing 7 / 66

8 A simple game Player A and Player B play a game. The probability that Player A wins each game is 0.6 and the probability that Player B wins each game is 0.4. They play the game N times. Each game is independent. Let X i = 0 if Player A wins game i X i = 1 if Player B wins game i Also assume there is an initial Game called Game 0 (X 0 ) Feng Li (SAM.CUFE.EDU.CN) Statistical Computing 8 / 66

9 Some simple questions What is the probability that Player A wins Game 1 ((X 1 = 0)) if If X 0 = 0 (Player A wins Game 0) If X 0 = 1 (Player B wins Game 0) What is the probability that Player A wins Game 2 ((X 2 = 0)) if If X 0 = 0 (Player A wins Game 0) If X 0 = 1 (Player B wins Game 0) Since each game is independent all answers are 0.6. Feng Li (SAM.CUFE.EDU.CN) Statistical Computing 9 / 66

10 A different game: A Markov chain Now assume that both players have a better chance of winning Game i + 1 if they already won Game i. Pr(X i+1 = 0 X i = 0) = 0.8 (9) Pr(X i+1 = 1 X i = 1) = 0.7 (10) Assume nothing other than game i has a direct effect on Game i + 1. This is called the Markov Property. Mathematically Pr(X i+1 X i, X i 1,..., X 1, X 0 ) = Pr(X i+1 X i ) (11) Feng Li (SAM.CUFE.EDU.CN) Statistical Computing 10 / 66

11 Markov Property Another way to define the Markov property is to notice that X i+1 and X i 1,..., X 0 are independent conditional on X i This may be a model for the stock market, all the valuable information about tomorrow s stock price is contained in today s price. This is related to the Efficient Market Hypothesis, a popular theory in finance. Now back to the simple game. Feng Li (SAM.CUFE.EDU.CN) Statistical Computing 11 / 66

12 Simulating from a Markov chain Now let s simulate a sequence X 1, X 2,..., X 100 from the Markov chain. Initialize at x 0 = 0. Then inside a loop Code the following using if. " 0 with probability 0.8 if X i = 0 then X i+1 = 1 with probability 0.2 " 0 with probability 0.3 if X i = 1 then X i+1 = 1 with probability 0.7 Try it Feng Li (SAM.CUFE.EDU.CN) Statistical Computing 12 / 66

13 Markov chain Sequence of X x i Feng Li (SAM.CUFE.EDU.CN) Statistical Computing 13 / 66

14 Simple questions again What is the probability that Player A wins the first game (i.e (X 1 = 0)) if If X 0 = 0 (Player A wins initial game) If X 0 = 1 (Player B wins initial game) The answers are 0.8 and 0.3. What is the probability that Player A wins the second game (X 2 = 0) if If X 0 = 0 (Player A wins initial game) If X 0 = 1 (Player B wins initial game) Feng Li (SAM.CUFE.EDU.CN) Statistical Computing 14 / 66

15 Solution Let X 0 = 0. Then Pr(X 2 = 0 X 0 = 0) = ÿ Pr(X 2 = 0, X 1 = x 1 X 0 = 0) x 1 =0,1 = ÿ Pr(X 2 = 0 X 1 = x 1, X 0 = 0)Pr(X 1 = x 1 X 0 = 0) x 1 =0,1 = ÿ Pr(X 2 = 0 X 1 = x 1 )Pr(X 1 = x 1 X 0 = 0) x 1 =0,1 = 0.8 ˆ ˆ 0.2 = 0.7 What if X 0 = 1? Feng Li (SAM.CUFE.EDU.CN) Statistical Computing 15 / 66

16 Recursion Notice that the distribution of X i depends on X 0 The sequence is no longer independent. How could you compute Pr(X n = 0 X 0 = 0) when n = 3, when n = 5, when n = 100? This is hard, but the Markov Property does make things simpler We can use a recursion to compute the probability that Player A wins any game. Feng Li (SAM.CUFE.EDU.CN) Statistical Computing 16 / 66

17 Recursion Note that Pr(X i = 0 X 0 = 0) = ÿ Pr(X i = 0, X i 1 = x i 1 X 0 = 0) x i 1 = ÿ x i 1 Pr(X i = 0 X i 1 = x i 1, X 0 = 0)Pr(X i 1 = x i 1 X 0 = 0) = ÿ x i 1 Pr(X i = 0 X i 1 = x i 1 )Pr(X i 1 = x i 1 X 0 = 0) We already applied this formula when i = 2. We can continue for i = 3, 4, 5,..., n Feng Li (SAM.CUFE.EDU.CN) Statistical Computing 17 / 66

18 Recursion Pr(X i = 0 X 0 = 0) = ÿ x i 1 Pr(X i = 0 X i 1 = x i 1 )Pr(X i 1 = x i 1 X 0 = 0) Start with Pr(X 1 = 0 X 0 = 0) Get Pr(X 1 = 1 X 0 = 0) Use these in formula with i = 2 Get Pr(X 2 = 0 X 0 = 0) Get Pr(X 2 = 1 X 0 = 0) Use these in formula with i = 3 Get Pr(X 3 = 0 X 0 = 0).... Feng Li (SAM.CUFE.EDU.CN) Statistical Computing 18 / 66

19 Matrix Form It is much easier to do this calculation in matrix form (especially when X is not binary). Let P be the transition matrix X i = 0 X i = 1 X i 1 = 0 Pr(X i = 0 X i 1 = 0) Pr(X i = 1 X i 1 = 0) X i 1 = 1 Pr(X i = 0 X i 1 = 1) Pr(X i = 1 X i 1 = 1) Feng Li (SAM.CUFE.EDU.CN) Statistical Computing 19 / 66

20 Matrix Form In our example: X i = 0 X i = 1 X i 1 = X i 1 = P = ( ) (12) Feng Li (SAM.CUFE.EDU.CN) Statistical Computing 20 / 66

21 Matrix Form Let π i be a 1 ˆ 2 row vector which denotes the probabilities of each player winning Game i conditional on the initial Game In our example if X 0 = 0 π i = (Pr(X i = 0 X 0 ), Pr(X i = 1 X 0 )) (13) In our example if X 0 = 1 π 1 = (0.8, 0.2) (14) π 1 = (0.3, 0.7) (15) Feng Li (SAM.CUFE.EDU.CN) Statistical Computing 21 / 66

22 Recursion in Matrix form The recursion formula is π i = π i 1 P (16) Therefore Now code this up in R. What is Pr(X n = 0 X 0 = 0) when n = 3 n = 5 n = 100? Do the same when X 0 = 1 π n = π 1 P ˆ P ˆ... ˆ P (17) Feng Li (SAM.CUFE.EDU.CN) Statistical Computing 22 / 66

23 Convergence? For n = 3 and n = 5, the starting point made a big difference. For n = 100 it did not make a big difference. Could this Markov chain be converging to something? Now write code to keep the values of π i for i = 1, 2,..., 100. Then plot the values of π i1 against i Feng Li (SAM.CUFE.EDU.CN) Statistical Computing 23 / 66

24 Convergence X 0 = 0 Pr(X i = 0) i X 0 = 1 Pr(X i = 0) i Feng Li (SAM.CUFE.EDU.CN) Statistical Computing 24 / 66

25 More Questions What is Pr(X 100 = 0 X 0 = 0)? What is Pr(X 100 = 0 X 0 = 1)? What is Pr(X 1000 = 0 X 0 = 0)? What is Pr(X 1000 = 0 X 0 = 1)? The answer to all of these is 0.6. The X do not converge. They keep changing from 0 to 1. The Markov chain however converges to a stationary distribution. Feng Li (SAM.CUFE.EDU.CN) Statistical Computing 25 / 66

26 Simulation with a Markov chain Go back to your code for generating a Markov chain and generate a chain with n = Exclude the first values of X i and keep the remaining values. How many X i = 0? How many X i = 1 We have discovered a new way to simulate from a distribution with Pr(X i = 0) = 0.6 and Pr(X i = 1) = 0.4 Feng Li (SAM.CUFE.EDU.CN) Statistical Computing 26 / 66

27 Markov Chains Sometimes two different Markov chains converge to the same stationary distribution. See what happens when ( ) P = (18) Sometimes Markov chains do not converge to a stationary distribution at all. Some Markov chains can get stuck in an absorbing state. For example what would the simple example look like if Pr(X i+1 = 0 X i = 0) = 1? Markov chains can be defined on continuous support as well, X i can be continuous. Feng Li (SAM.CUFE.EDU.CN) Statistical Computing 27 / 66

28 Some important points This is a very complicated way to generate from a simple distribution. For the binary example the direct method would be better. However for other examples, either the direct method or accept/reject algorithm do not work. In these cases we can construct a Markov chain that has a stationary distribution that is our target distribution. All we need is the kernel of the density function, and an algorithm called the Metropolis Algorithm Feng Li (SAM.CUFE.EDU.CN) Statistical Computing 28 / 66

29 Normalizing Constant and Kernel What are the normalizing constant and kernel of the Beta density? Beta(x; a, b) = Γ(a + b) (Γ(a)Γ(b)) xa 1 (1 x) b 1 (19) Feng Li (SAM.CUFE.EDU.CN) Statistical Computing 29 / 66

30 The Metropolis algorithm The Metropolis algorithm was developed in a 1953 paper by Metropolis, Rosenbluth, Rosenbluth, Teller and Teller. The aim is to simulate x p(x) where p(x) is called the target density. We will need a proposal density q(x [old] Ñ x [new] ) For example one choice of q is This is called a Random Walk proposal x [new] N(x [old], 1) (20) Feng Li (SAM.CUFE.EDU.CN) Statistical Computing 30 / 66

31 Symmetric proposal An important property of q in the Metropolis algorithm is symmetry of the proposal q(x [old] Ñ x [new] ) = q(x [new] Ñ x [old] ) (21) Later we will not need this assumption Can you confirm this is true for x [new] N(x [old], 1)? Can you simulate from this random walk (use x 0 = 0 as a starting value)? Feng Li (SAM.CUFE.EDU.CN) Statistical Computing 31 / 66

32 Proof of symmetry of random walk The proposal " ) * 2 q(x [old] Ñ x [new] ) = (2π) 1/2 exp 1 (x [new] x [old] 2 " [ )] * 2 = (2π) 1/2 exp 1 1 (x [new] x [old] 2 " ) * 2 = (2π) 1/2 exp 1 (x [old] x [new] 2 (22) (23) (24) = q(x [new] Ñ x [old] ) (25) Feng Li (SAM.CUFE.EDU.CN) Statistical Computing 32 / 66

33 Accept and reject By itself the random walk will not converge to anything. To make sure this Markov chain converges to our target, we need to include the following. At step i + 1 set x [old] = x [i]. Generate x [new] N(x [old], 1) and compute ( ) α = min 1, p(x[new] ) p(x [old] ) Then Set x [i+1] to x [new] with probability α (accept) Set x [i+1] to x [old] with probability 1 α (reject) (26) Feng Li (SAM.CUFE.EDU.CN) Statistical Computing 33 / 66

34 Code it up Use the Metropolis algorithm with a random walk proposal to simulate a sample from the standard t distribution with 5 df. The target density is p(x) = ] 3 [1 + x2 (27) 5 Simulate a Markov chain with iterates using the random walk as a proposal. The first iterates are the burn-in and will be left out because the Markov chain may not have converged yet. Use x 0 = 0 as a starting value Feng Li (SAM.CUFE.EDU.CN) Statistical Computing 34 / 66

35 Some diagnostics - Convergence There a few diagnostics we can use to investigate the behaviour of the chain One is a trace plot (including burn in), which is simply a line plot of the iterates. Plot this for your Markov chain Another diagnostic is the Geweke diagnostic which can be found in the R package coda. The Geweke diagnostic tests the equality of the means of two different parts of the chain (excluding burn-in). The test statistic has a standard normal distribution. Rejecting this test is evidence that the chain has not converged Feng Li (SAM.CUFE.EDU.CN) Statistical Computing 35 / 66

36 Trace Plot Initial Value =0, Proposal S.D. =1 x e+00 2e+04 4e+04 6e+04 8e+04 1e+05 Feng Li (SAM.CUFE.EDU.CN) Statistical Computing 36 / 66

37 The effect of starting value Now rerun the code with a starting value of X 0 = 100 Does the chain still converge? Does it converge quicker or slower? Feng Li (SAM.CUFE.EDU.CN) Statistical Computing 37 / 66

38 Trace Plot Initial Value =0, Proposal S.D. =1 x e+00 2e+04 4e+04 6e+04 8e+04 1e+05 Feng Li (SAM.CUFE.EDU.CN) Statistical Computing 38 / 66

39 Trace Plot Initial Value =100, Proposal S.D. =1 x e+00 2e+04 4e+04 6e+04 8e+04 1e+05 Feng Li (SAM.CUFE.EDU.CN) Statistical Computing 39 / 66

40 The effect of proposal variance Keep the starting value of X 0 = 100 No change the standard deviation of the proposal to 3. Does the chain still converge? Does it converge quicker or slower? Feng Li (SAM.CUFE.EDU.CN) Statistical Computing 40 / 66

41 Trace Plot Initial Value =100, Proposal S.D. =1 x e+00 2e+04 4e+04 6e+04 8e+04 1e+05 Feng Li (SAM.CUFE.EDU.CN) Statistical Computing 41 / 66

42 Trace Plot Initial Value =100, Proposal S.D. =3 x e+00 2e+04 4e+04 6e+04 8e+04 1e+05 Feng Li (SAM.CUFE.EDU.CN) Statistical Computing 42 / 66

43 Huge proposal variance Maybe you think the best strategy is to choose a huge standard deviation. Try to use a proposal standard deviation of 100. Plot a trace plot of the chain. The plot is rejecting many iterates. This must be inefficient Change your code to compute the percentage of times a new iterate is accepted (excluding burn in). Use x 0 = as an initial value. What is the acceptance rate when the proposal standard deviation is 1? What is the acceptance rate when the proposal standard deviation is 100? Feng Li (SAM.CUFE.EDU.CN) Statistical Computing 43 / 66

44 Trace Plot Initial Value =100, Proposal S.D. =100 x e+00 2e+04 4e+04 6e+04 8e+04 1e+05 Feng Li (SAM.CUFE.EDU.CN) Statistical Computing 44 / 66

45 Acceptance Rate If the proposal variance is too high Values will be proposed that are too far into the tails of the stationary distribution The Metropolis algorithm will mostly reject these values. The sample will still come from the correct target distribution but this is a very inefficient way to sample. What happens if a very small proposal variance is used. Try a proposal variance of What is the acceptance rate? Feng Li (SAM.CUFE.EDU.CN) Statistical Computing 45 / 66

46 Trace Plot Initial Value =0, Proposal S.D. =0.001 x e+00 2e+04 4e+04 6e+04 8e+04 1e+05 Feng Li (SAM.CUFE.EDU.CN) Statistical Computing 46 / 66

47 Acceptance Rate The acceptance rate is almost 1. However is this a good proposal? The jumps made by this proposal are too small, and do not sample enough iterates from the tails of the distribution. If it runs long enough the Markov chain will provide a sample from the target distribution. However, it is very inefficient. Feng Li (SAM.CUFE.EDU.CN) Statistical Computing 47 / 66

48 Random Walk proposal For a random walk proposal it is not good to have an acceptance rate that is too high or too low. What exactly is too high and too low? It depends on many things including the target and proposal. A rough rule is to aim for an acceptance rate between 20% and 70% If your acceptance rate is outside this range the proposal variance can be doubled or halved There are better (but more complicated) ways to do this. Feng Li (SAM.CUFE.EDU.CN) Statistical Computing 48 / 66

49 Monte Carlo Error Now that there is a sample. X [1], X [2],..., X [M] p(x). What can it be used for? We can estimate the expected value E(X) This can be done by taking: E(X) «1 M Note we use «instead of =. There is some error since we are estimating E(X) based on a sample. Mÿ i=1 X [i] Luckily we can make this smaller by generating a bigger sample. We call this Monte Carlo error. Feng Li (SAM.CUFE.EDU.CN) Statistical Computing 49 / 66

50 Measuring Monte Carlo Error One way to measure Monte Carlo Error is the variance of the sample mean. ( ) ( ) 1 Mÿ Var X [i] = 1 M M M 2 Var ÿ X [i] i=1 i=1 = 1 Mÿ M 2 Var(X [i] ) i=1 = Var(X) M A sample from a Markov chain is correlated Feng Li (SAM.CUFE.EDU.CN) Statistical Computing 50 / 66

51 Measuring Monte Carlo Error One way to measure Monte Carlo Error is the variance of the sample mean. ( ) ( ) 1 Mÿ Var X [i] = 1 M M M 2 Var ÿ X [i] i=1 i=1 = 1 Mÿ M 2 Var(X [i] ) i=1 = Var(X) M A sample from a Markov chain is correlated Feng Li (SAM.CUFE.EDU.CN) Statistical Computing 50 / 66

52 Measuring Monte Carlo Error One way to measure Monte Carlo Error is the variance of the sample mean. ( ) ( ) 1 Mÿ Var X [i] = 1 M M M 2 Var ÿ X [i] i=1 i=1 = 1 ÿ M M 2 Var(X [i] ) + 2 ÿ M ÿ M 2 cov(x [i], X [j] ) i=1 = Var(X) M + 2 M 2 i=1 jąi Mÿ ÿ cov(x [i], X [j] ) i=1 jąi A sample from a Markov chain is correlated Feng Li (SAM.CUFE.EDU.CN) Statistical Computing 50 / 66

53 Monte Carlo efficiency It is better to have lower correlation in the Markov chain. The efficiency of the chain can be measured using the effective sample size. The effective sample size can be computed using the function effectivesize in the R Package coda. Obtain an effective sample size for your sample (excluding burn in) where Proposal S.D. =1 Proposal S.D. =5 My answers were about 6000 and and yours should be close to that Feng Li (SAM.CUFE.EDU.CN) Statistical Computing 51 / 66

54 Interpret Effective Sample Size What does it mean to say a Monte Carlo with a sample size of has an effective sample size (ESS) of just 6000? The sampling error of a correlated sample of is equal to the sampling error of an independent sample of Mathematically ( ) 1 Mÿ Var X [i] M i=1 = Var(X) M + 2 M 2 Mÿ Mÿ i=1 jąi cov(x [i], X [j] ) = Var(X) M eff It is a useful diagnostic for comparing two different proposal variances. A higher ESS implies a more efficient scheme. Feng Li (SAM.CUFE.EDU.CN) Statistical Computing 52 / 66

55 Non-Symmetric proposal In 1970, Hastings proposed an extension to the Metropolis Hastings algorithm. This allows for the case when q(x [old] Ñ x [new] ) q(x [new] Ñ x [old] ) (28) The only thing that changes is the acceptance probability ( α = min 1, p(x[new] )q(x [new] Ñ x [old] ) ) p(x [old] )q(x [old] Ñ x [new] ) This is called the Metropolis-Hastings algorithm (29) Feng Li (SAM.CUFE.EDU.CN) Statistical Computing 53 / 66

56 An interesting proposal Suppose we use the proposal: x new N(0, a (5/3)) (30) What is q(x [old] Ñ x [new] )? It is q(x [new] ) where q(.) is the density of a N(0, a (5/3)). Is this symmetric? No, since generally q(x [new] ) q(x [old] ) Feng Li (SAM.CUFE.EDU.CN) Statistical Computing 54 / 66

57 Metropolis Hastings Code this where p(.) is the standard t density with 5 d.f, and q(.) is normal with mean 0 and standard deviation 5/3. Inside a loop Generate x [new] N(0, a (5/3)) Set x old = x [i] and compute ( ) α = min 1, p(x[new] )q(x [old] ) (31) p(x [old] )q(x [new] ) Try it Set x [i+1] to x [new] with probability α (accept) Set x [i+1] to x [old] with probability 1 α (reject) Feng Li (SAM.CUFE.EDU.CN) Statistical Computing 55 / 66

58 Comparison The Effective Sample Size of this proposal is about much higher than the best random walk proposal. Why does it work so well? The standard t distribution with 5 df has a mean of 0 and a standard deviation of a (5/3) So the N(0, a (5/3)) is a good approximation to the standard student t with 5 df. Feng Li (SAM.CUFE.EDU.CN) Statistical Computing 56 / 66

59 Laplace Approximation Using a Taylor expansion of lnp(x) around the point a lnp(x) «lnp(a) + Blnp(x) ˇ (x a) + 1 B 2 lnp(x) ˇ Bx ˇx=a 2 Bx 2 (x a) 2 ˇx=a Let a be the point that maximises lnp(x) and let ( B 2 lnp(x) b = ˇ Bx 2 ˇx=a ) 1 (32) Feng Li (SAM.CUFE.EDU.CN) Statistical Computing 57 / 66

60 The approximation is lnp(x) «lnp(a) 1 (x a)2 2b Taking exponential of both sides ] [ (x a)2 p(x) «k ˆ exp 2b Any distribution can be approximated by a normal distribution with mean a and variance b where a and b values can be found numerically if needed. Feng Li (SAM.CUFE.EDU.CN) Statistical Computing 58 / 66

61 Exercise: Generating from skew normal The density of the skew normal is p(x) = 2φ(x)Φ(δx) (33) where φ(x) is the density of the standard normal Φ(x) is the distribution of the standard normal. Using a combination of optim, dnorm and pnorm find the Laplace approximation of the skew normal when δ = 3 Use it to generate a sample from the skew normal distribution using the Metropolis Hastings algorithm. Feng Li (SAM.CUFE.EDU.CN) Statistical Computing 59 / 66

62 Indirect method v Metropolis Hastings Some similarities are: Both require a proposal Both are more efficient when the proposal is a good approximation to the target. Both involve some form of accepting/rejecting Some differences are: Why? Indirect method produces an independent sample, MH samples are correlated. Indirect method requires p(x)/q(x) to be finite for all x. MH works better when x is a (high-dimensional vector). Feng Li (SAM.CUFE.EDU.CN) Statistical Computing 60 / 66

63 Multiple variables Suppose we now want to sample from a bivariate distribution p(x, z) The ideas involved in this section work for more than two variables. It is possible to do a 2-dimensional random walk proposal. However as the number of variables goes up the acceptance rate becomes lower. Also the Laplace approximation does not work as well in high dimensions. Indirect methods of simulation suffer from the same problem. We need a way to break the problem down. Feng Li (SAM.CUFE.EDU.CN) Statistical Computing 61 / 66

64 Method of composition Markov chain methods, allow us to break multivariate distributions down. If it is easy to generate from p(x) then the best way is Method of composition. Generate x [i] p(x) z [i] p(z x = x [i] ) Sometimes p(x) is difficult to get Feng Li (SAM.CUFE.EDU.CN) Statistical Computing 62 / 66

65 Gibbs Sampler If it is easy to simulate from the conditional distribution f(x z) then that can be used as a proposal What is the acceptance ratio? α = = = 1 ( 1, p(xnew, z)p(x old ) z) p(x old, z)p(x new z) ( 1, p(xnew z)p(z)p(x old ) z) p(x old z)p(z)p(x new z) Feng Li (SAM.CUFE.EDU.CN) Statistical Computing 63 / 66

66 Gibbs Sampler This gives the Gibbs Sampler Generate x [i+1] p(x [i+1] z [i] ) Generate z [i+1] p(z [i+1] x [i+1] ) Repeat x and z can be swapped around. It works for more than two variables. Always make sure the conditioning variables are at the current state. Feng Li (SAM.CUFE.EDU.CN) Statistical Computing 64 / 66

67 Metropolis within Gibbs Even if the individual conditional distributions are not easy to simulate from, Metropolis Hastings can be used within each Gibbs step. This works very well because it breaks down a multivariate problem into smaller univariate problems. We will practice some of these algorithms in the context of Bayesian Inference Feng Li (SAM.CUFE.EDU.CN) Statistical Computing 65 / 66

68 Summary You should be familiar with a Markov chain You should understand this can have a stationary distribution You should have a basic understanding of the Metropolis Hastings and the special cases Random Walk Metropolis Laplace approximation Gibbs Sampler Feng Li (SAM.CUFE.EDU.CN) Statistical Computing 66 / 66